Google DeepMind argues AI safeguards need both lab rules and local judgement

Google DeepMind’s Manish Gupta argues that AI safeguards should combine rigorous checks by developers with rules tailored to the settings where systems are used. His essay also points to safety work and partnerships the company says it is pursuing in India, where local language and domain needs are part of the challenge.

Google DeepMind Watch analysis

What happened

Gupta, a Senior Director of Research at Google DeepMind, sets out the company’s view of AI safety in an essay published by India Today. He argues that developers should handle baseline safety, evaluations and documentation, while deployers and sector regulators set context-specific safeguards for fields such as emergency triage and banking.

The essay describes internal red-teaming and external evaluations, and names work including the Hindi-language AILuminate Safety Benchmark with CeRAI at IIT Madras and MLCommons. It also says AIIMS teams are adapting MedGemma for local clinical needs, and that Project Vaani has open-sourced more than 31,000 hours of speech data across more than 100 Indic languages. These are claims and examples presented by Gupta, not independent assessments of their results. Read Gupta’s essay in India Today.

Why it matters

The division of responsibility is the useful part of the argument. A model developer can test a system and document its limits, but cannot set every safeguard for every hospital, bank or public service that later uses it. That leaves a practical question for regulators and deployers: how to add protections for local needs without simply duplicating the work done upstream.

The India examples also put language access and local expertise on the safety agenda, rather than treating them as optional polish after a model is built.

Our read

Gupta offers a clear framework, but this is also a company executive making the case for his organisation’s approach. The proposal is worth debating; the examples are worth testing against published methods, independent evaluations and evidence of outcomes. “Safety” is a good ambition. The interesting work is showing who is responsible for what, and how anyone outside the lab can tell whether it is working.

What to watch

• Whether the proposed division of responsibility becomes specific, enforceable guidance.
• What methods and results are published for the named safety benchmarks and tools.
• Whether local-language and domain partnerships produce independently assessed improvements.

Discussion spark: Should AI developers be responsible for safety across every use of their models, or should deployers and regulators carry the main burden of adapting safeguards to local contexts?

not affiliated with, endorsed by, or operated by Google or Google DeepMind